Frontiers in Marine Science (Apr 2024)
Deep learning–assisted biodiversity assessment in deep-sea benthic megafauna communities: a case study in the context of polymetallic nodule mining
Abstract
IntroductionTechnological developments have facilitated the collection of large amounts of imagery from isolated deep-sea ecosystems such as abyssal nodule fields. Application of imagery as a monitoring tool in these areas of interest for deep-sea exploitation is extremely valuable. However, in order to collect a comprehensive number of species observations, thousands of images need to be analysed, especially if a high diversity is combined with low abundances such is the case in the abyssal nodule fields. As the visual interpretation of large volumes of imagery and the manual extraction of quantitative information is time-consuming and error-prone, computational detection tools may play a key role to lessen this burden. Yet, there is still no established workflow for efficient marine image analysis using deep learning–based computer vision systems for the task of fauna detection and classification.MethodsIn this case study, a dataset of 2100 images from the deep-sea polymetallic nodule fields of the eastern Clarion-Clipperton Fracture zone from the SO268 expedition (2019) was selected to investigate the potential of machine learning–assisted marine image annotation workflows. The Machine Learning Assisted Image Annotation method (MAIA), provided by the BIIGLE system, was applied to different set-ups trained with manually annotated fauna data. The results computed with the different set-ups were compared to those obtained by trained marine biologists regarding accuracy (i.e. recall and precision) and time.ResultsOur results show that MAIA can be applied for a general object (i.e. species) detection with satisfactory accuracy (90.1% recall and 13.4% precision), when considered as one intermediate step in a comprehensive annotation workflow. We also investigated the performance for different volumes of training data, MAIA performance tuned for individual morphological groups and the impact of sediment coverage in the training data.DiscussionWe conclude that: a) steps must be taken to enable computer vision scientists to access more image data from the CCZ to improve the system’s performance and b) computational species detection in combination with a posteriori filtering by marine biologists has a higher efficiency than fully manual analyses.
Keywords
